ai.onnx.DynamicQuantizeLinear

ai.onnx · standard ONNX operator · ONNX opset ≥ 11

Description

Computes a per-tensor scale and zero point from the range of floating-point input x, extending the range to include zero, then quantizes each value to uint8 as saturate(round(x / y_scale) + y_zero_point). Uses round-to-nearest-even and clamps results to [0, 255].

See the ONNX DynamicQuantizeLinear spec for the reference semantics.

Inputs

Outputs

Type constraints

Implementation variants

One implementation is selected per call from the device capabilities, the request shapes and the dtypes; these notes say what each one covers.

• single_invocation — Uses one invocation to find the range and quantize the tensor, avoiding partial buffers for small inputs. It also provides the fallback when the parallel reduction cannot satisfy device limits.
• parallel_subgroup_reduce_vec4 — Reduces independent input blocks to min/max partials, combines them, and quantizes in a separate pass. The family uses packed reads when the input length is vec4-aligned.
• parallel_subgroup_reduce — Reduces independent input blocks to min/max partials, combines them, and quantizes in a separate pass. The family uses packed reads when the input length is vec4-aligned.
• grid_stride_reduce_vec4 — Caps the number of min/max partials and grid-strides each workgroup across the input. This bounds scratch size and finalization work for large tensors.
• grid_stride_reduce — Caps the number of min/max partials and grid-strides each workgroup across the input. This bounds scratch size and finalization work for large tensors.

Files

• metadata.json — kernel metadata (id, digests, per-variant templates, provenance)
• manifest.json — the op contract (source of truth)
• test.json — correctness cases
• bench.json — benchmark + tuning cases
• dynamic-quantize-linear-quantize.wgsl.jinja
• dynamic-quantize-linear-reduce.wgsl.jinja
• dynamic-quantize-linear.wgsl.jinja

Use with @huggingface/kernels

npm install --save-exact @huggingface/kernels@0.0.1-preview.2

Required output shapes and logical data types are inferred from the supplied inputs and attributes; result tensors are allocated automatically.

The version: 1 option selects the published kernel contract; it is independent of any operator opset, contrib since_version, or model version. It follows the v1 branch as fixes land. To pin exact artifact bytes, pass a 40-character commit revision instead of version.

Replace each *Data placeholder with a typed array containing the corresponding input data.

import { getKernel } from "@huggingface/kernels";

const kernel = await getKernel("webgpu-kernels/ai.onnx.DynamicQuantizeLinear", { version: 1 });
const { y, y_scale, y_zero_point } = await kernel({ x: { data: xData, shape: [1] } });